Prediction of wall deflection induced by braced excavation in spatially variable soils via convolutional neural network

Abstract

Recently, the random field finite element method (RF-FEM) has attracted significantly increasing attention in the field of geotechnical engineering, especially for the purpose of analyzing the response of geotechnical systems due to the inherent variability of physical and mechanical properties. However, the method requires repeated finite element calculations based on a mass of sampling processes, making the computing effort expensive. The surrogate model is one of the techniques commonly adopted to alleviate the computational burden. In addition, some architectures of deep learning surrogate models are so unique that it is difficult to transfer to similar cases and to be familiar and reproducible by readers. In this study, we propose a convolutional neural network (CNN) surrogate model based on classical architecture, VGG6, to perform random field finite element analyses (RF-FEM). We pre-process the tabular data generated by the random filed method into an image-like format as input data. The VGG6 is used as a surrogate model to replace the original RF-FEM simulations for all subsequent calculations. The applications of the proposed method to assess wall deflection of braced excavation in clays with randomly varying cohesion cu and the friction angle φ are illustrated and compared in different cases. The excellent agreement between the VGG6 outputs and the FEM predictions demonstrated the promising potential of using VGG6s as a surrogate model for reliability analysis in spatially variable soils. Moreover, the model saves a lot of computing time and computing power and fully proves the generalization performance of the model under various cases.